MCP Server

Overview

Retrace provides a hosted Model Context Protocol (MCP) server that lets AI coding agents search traces, recall memories, and inspect past executions directly from their IDE. This bridges your observability data with your development workflow — agents can learn from previous runs without leaving the editor.

Setup

Connect your AI coding agent to Retrace's hosted MCP endpoint:

Cursor / Windsurf (MCP config):

{
  "mcpServers": {
    "retrace": {
      "url": "https://api.retraceai.tech/mcp",
      "headers": {
        "Authorization": "Bearer rt_live_..."
      }
    }
  }
}

Claude Code:

claude mcp add --transport http retrace https://api.retraceai.tech/mcp --header "Authorization: Bearer rt_live_..."

Codex:

[mcp_servers.retrace]
url = "https://api.retraceai.tech/mcp"
http_headers = { Authorization = "Bearer rt_live_..." }

Replace rt_live_... with your Retrace API key from Settings.

Available Tools

Read Tools

Mutation Tools

Pagination

All list tools support offset for pagination. Responses include:

{
  "results": [...],
  "offset": 0,
  "limit": 10,
  "hasMore": true
}

To page through results: increment offset by limit until hasMore is false.

Example Prompts

Once configured, ask your AI coding agent natural-language questions:

• "What traces failed recently?" → calls list_traces with status: "failed"
• "Search for spans where the agent called web_search" → calls search_spans
• "Remember that the user prefers TypeScript" → calls store_memory
• "What do we know about the API rate limits?" → calls recall_memories
• "Show me the demo tape" → calls get_tape
• "Fork that trace from the second LLM call with a different prompt" → calls run_fork
• "Publish this trace as a shareable tape" → calls publish_tape
• "Create a new trace for my test run" → calls create_trace

Use Cases

Debugging — When an agent run fails, ask your coding agent to pull the trace and inspect which span errored. It can read inputs, outputs, and error messages without you opening the dashboard.

Knowledge persistence — Store architectural decisions, user preferences, or project conventions as memories. Your coding agent recalls them in future sessions automatically.

Learning from history — Search past successful traces to understand how a workflow executed, then replicate or improve the pattern.

Troubleshooting